Research Of Key Algorithms On Change Detection Of Land Cover Using Multi-Temporal Polarimetric SAR Imagery

With the sustained and rapid development of the global economy as well as the transformation of human beings into the natural environment,the changing of the surface land-cover type occurs frequently.A quick and accurate measurement of the information of different land-cover types has higher value in many applications,such as land resources survey,natural disaster monitoring,etc.The change detection of land cover using multi-temporal remote sensing imagery can overcome the shortcomings of field survey.More importantly,it has become a primary mean to study the change detection of land cover in regional and global scale.Polarimetric Synthetic Aperture Radar(PolSAR),working on the active imaging mode,can not only reduce the influence of optical sensor,which is susceptible to weather and other factors,but also obtain more detailed scatter information by transmitting and receiving electromagnetic waves of different polarizations when compared with single-polarization SAR.As a result,it provides data support of land-cover type's change with high accuracy.The rapid development of PolSAR technology makes it possible to collect large volume of PolSAR images at different time of the same area.It has become an urgent problem to comprehensively utilize multi-temporal polarimetric SAR imagery to extract the dynamic change information of land-cover types accurately.Therefore,this thesis summarizes the methods of change detection for land cover based on multi-temporal PolSAR images at present.As the multi-temporal PolSAR images can provide rich information of scattering characteristics and statistical distribution,this thesis takes the multi-temporal Radarsat-2 and Gaofen-3(GF-3)images as the main data source to explore the key algorithms on change detection of land cover by using statistical learning and machine learning technology.The main content of dissertation are as follows:(1)In order to solve the problem of the pixel-based unsupervised change detection method,which is susceptible speckle noise,high false alarm rate,and the poor performance of traditional threshold algorithm,a novel frame of unsupervised change detection using bi-temporal PolSAR images was proposed in this research.The difference image was segmented by using the generalized statistical region merging(GSRM)algorithm,which can enhance the homogeneity of the difference image in the‘suspected'change region.The Gaussian mixture model was improved by introducing the Elbow algorithm,which can fit the probability density function of difference image well and effectively distinguish the change and non-change categories.The experimental results indicate that the novel frame can efficiently improve the accuracy of unsupervised change detection.(2)Regarding that the existing unsupervised change detection methods were mainly applied to the analysis of bi-temporal PolSAR images and the time-series PolSAR images cannot be utilized in an optimal manner,this research introduced the Omnibus and R_j statistical hypothesis testing algorithm and proposed a novel frame of unsupervised change detection using time-series PolSAR images.The experimental results indicate that the overall change of the study area and change information at any time interval can be better extracted,the time point of changes in the region can be accurately identified,and the information of gradual,continuous change can be detected.In addition,the efficiency of time-series unsupervised change detection can be increased.(3)Considering that the supervised change detection methods using bi-temporal PolSAR images requires a large amount of labeled samples,this research employed the modified breaking ties(MBT)algorithm to obtain a small number of the most informative training samples.In order to cope with the problem that the change detection results are susceptible to the influence of speckle noise,this research introduced the object-oriented method to suppresses the influence of speckle noise.The experimental results show that the novel supervised change detection can reduce the labeling cost of training sample and improve the accuracy of change detection when compared with the traditional methods.(4)By combining with the Omnibus statistical hypothesis testing algorithm,active learning algorithm,and transfer learning algorithm,a supervised change detection using multi-temporal PolSAR images was proposed in this paper.The novel supervised change detection can solve the problems of existing methods,which do not make full use of time dimension information,and require a large number of training samples with high-quality.The experimental results indicate that the proposed method only requires small amounts of training samples within one PolSAR image and realizes the information extraction of dynamic change for land cover type.Moreover,it reduced the dependence of training samples for each PolSAR image and improve the efficiency of category dynamic change information.